Node Representation Learning Based Cross Network Node Correlation

Cross-network node correlation is to associates multiple accounts belonging to the same entity user in different online social networks(OSN).As one of the core tasks of data mining and network security on social websites,that has become a hot research filed nowadays.However,owing to the heterogeneous nature of the platform or some uncertain personal reasons,the same user may perform differently on different OSN platforms,which brings many challenges accordingly.Limited by the falsification of attributes and the time complexity of processing user-generated content,the overall performance of existing methods for cross-network node correlation is not ideal.This thesis mainly studies a cross-network node correlation method based on node representation of online social network.The friend relationship and user properties are extracted from the OSN website account,and then random walk on relational network is committed to learn the potential representation of the user,so that multiple social accounts belonging to the same entity user on different OSN platforms are associated.The main contents of the thesis are as follows:(1)proposes an easy-to-expand node representation method of comprehensive information.Based on friend relationship,personal profile,and hometown,characterize users' feature of network structure and attributes,through user-user walk and user-attribute walk.The learned vector representation portrays users' characteristics roundly and improves the accuracy of account correlation.(2)proposes a cross large-scale network node correlation algorithm.According to the corresponding vector representation of the anchor node,computes the transformation matrix to transform one vector space into the other.By this way,the associated user or candidate set can be found based on the calculation of vector similarity.Compared with the common matrix decomposition method,the accuracy,MAP and F-score of this method are improved.Besides,the correlation method based on network fusion is summarized,experiments comparison and application scene analysis are carried out.Finally,a self-learning iteration framework is described,and the feasibility is verified by the experiments.In general,we mainly study how to use social network structure and user's attributes to embed users in OSN into vector space,and associate cross-network nodes accordingly.Based on data sets of multiple single-network with different scale and a real-world associated-network,amounts of experimental data is analyzed in detail and conclusions are drawn.Compared with matrix decomposition and auto-encoders,the node representation of comprehensive information and spatial transformation proposed in this paper can improve the accuracy of cross network node correlation.